1.  Some units / scaling mismatch elsewhere (slip, ai, camber, Bxa, or Shxa) or a sign/duplicate shift being applied.
1b. Needing Cxa = 0.1 almost always means the argument inside atan() is 10× larger than the MF parameters were tuned for.

2. Camber (yi) units or sign //(In theory not the sign due you double checked, but what about the units?)
If camber yi is in degrees but the RBX3 coefficient assumes radians (or vice versa) you change Bxa unexpectedly.

3. dFz normalization mismatch leading to Exa outside intended range; large Exa can change curvature and effective response.

4. Shift Shxa applied twice or with wrong sign.
If you already added the shift when computing ai earlier, and then do As = ai + Shxa again, As will be too big.

5. Check G0 normalization term, an error here could drastically alter the effective shape of the curve and force you to change Cxa.